A TensorFlow loss function for binary classification based on an approximation of the normalized Wilcoxon-Mann-Whitney (WMW) statistic.
The normalized WMW statistic can be shown to be equal to the AUC-ROC. However, it is a step function so it is not differentiable. The normalized WCW statistic can be approximated with a smooth, differentiable function which makes the approximated version a near ideal loss function for optimizing for the AUC-ROC metric.
The loss function has two parameters, gamma and p, which are recommended to be kept between 0.1 to 0.7 and at 2 or 3, respectively.